Risk appetite is how much uncertainty the organisation will accept to get value from AI sooner. It decides which tools people may use and what goes into them, and some risk comes with the first person using AI at work.
The five positions differ most in when, and by whom, AI's output is checked: before anything is used, before work reaches a client, by the person who did the work, after the fact, or once a problem is reported.
Safe firstLink to this section
People may use AI only on approved tools, not on work involving client or personal information, and a person checks everything AI produces. The organisation gets the fewest incidents on the tools it can see, each approved after checking who sees what goes in, where it is kept, and whether the vendor may use it to improve its models. Most of the value that needs client or personal information is given up, and where checking takes nearly as long as the work, little is saved.
Easy to miss. With Safe first the risk is still there, in personal accounts the organisation cannot see, and it grows as people need more than the approved tools can do. In a global survey of 31,000 knowledge workers, 78% of AI users said they were bringing their own AI tools to work (Microsoft and LinkedIn, 2024 Work Trend Index Annual Report, 2024).
When to choose it. It makes sense where regulators or clients watch closely, and the approved tools are good enough, and added quickly enough, that people prefer them.
CautiousLink to this section
People use approved tools on internal work, such as drafting and research, without a check first. Anything reaching a client is checked by a person, and each tool has clear policies on which types of information can go in. Client work gains less than internal work, because each piece waits for a check.
Easy to miss. Much internal work contains client and personal information: meeting notes, a pricing draft, a message about a member of staff. The protection comes from the policies on each tool.
When to choose it. Choose it where most of the value is in internal work, and the policies are plain enough to apply to an ordinary email.
BalancedLink to this section
Most people use AI on real work, client work included, and check their own output. Policies are set by type of information rather than by tool, so new tools can be used within them. Most teams get value from it, and occasional errors get past the first check, so reviewing AI output becomes everyone's job.
Easy to miss. Balanced asks a great deal of judgement: each person decides what is safe to put into a tool, and people read the same policy differently, some teams as permission and others as prohibition.
When to choose it. It is best where the policies come with examples from the organisation's own work, and someone checks how each team applies them.
OpenLink to this section
People are encouraged to try new tools on real work, client work included. A few simple policies cover the things never allowed, such as client or personal information in an unapproved tool. Value comes faster, with more failed experiments, more errors reaching clients and more leadership time on them.
Easy to miss. Permission is taken up mostly by the already keen, increasing the difference between confident and unsure staff in results and morale.
When to choose it. It suits organisations where the less confident get deliberate support, such as time with a colleague who uses AI well, and someone has time to check after the fact.
Willing to take risksLink to this section
People find their own uses for AI and share what works, with very few policies and a fast way to report and fix problems. It suits confident self-starters and unsettles people who want clear policies. Value comes from the widest range of uses, and incidents are expected, which the organisation has to be able to recover from. Privacy law and client contracts still apply, so personal information goes only into tools that have been checked, and an incident involving it may have to be reported.
Easy to miss. Beyond data leaking, there is the risk of a wrong answer that looks right, which nobody catches because the person using the tool doesn't know the subject well enough to notice.
When to choose it. It is worth choosing where people experimenting know their field well, problems are reported quickly and without blame, privacy law is met, and insurers and clients accept it.
How to chooseLink to this section
- Which position describes the organisation today, in practice rather than on paper? A position below that takes tools away from people.
- What would a serious incident look like, could the organisation recover from it, and what do regulators, insurers, clients and contracts require?
- Is the value in internal work, or in work clients see?
- Who will do the checking, do they have the time, and do they know the subject well enough to spot a wrong answer that looks right?
- Does everyone mean the same thing? Ask each leader for one ordinary task that would be allowed and one that would not.
Tensions with other decisionsLink to this section
- Growth as the aim. Growth means using AI in work clients see, which is mostly off limits under Safe first and allowed with a check under Cautious.
- A long off-limits list. Open and Willing to take risks rely on a few clear limits. Five or six limits send a second message, and people tend to follow the written one.
- A fast pace, or the slowest. Safe first or Cautious at Quick or faster, or Safe first at Carefully, over time, tends to send people to their own accounts.
- A slow pace. Open or Willing to take risks at a slow pace suggests something else, such as money or confidence, is limiting the organisation.
- The fastest pace. Willing to take risks at As fast as we can takes a great deal of leadership time.
- Little money. Open or Willing to take risks at Quick or faster needs new tools and someone to sort out problems, both paid for centrally.
- Teams deciding. With Safe first or Cautious, teams choosing for themselves keep approved lists that become more different over time.
- Little time to learn. From Balanced upwards, judgement depends on whatever each person has learnt.
When to revisit itLink to this section
- After an incident or near miss: was the position wrong, or how it was applied?
- When regulators, insurers, clients or the aims change.
- On a date set now, because the tools change quickly.
For support with this decision, see Getting outside help.
Something missing or out of date here? Tell us.